Dynamic Stream Selection Network for Subject-Independent EEG-Based Emotion Recognition

Abstract

Due to severe cross-subject data variations in electroencephalogram (EEG) signals, the issue of subject-independent EEG-based emotion recognition remains challenging till today. To cope with this challenge, we propose a novel and effective dynamic stream selection network (DSSN), which can adaptively adjust its structure according to the characteristics of signal data from different individuals for this issue. DSSN consists of a tri-stream structure and a dynamic selection network. The tri-stream structure takes charge of extracting the spatial, the temporal, and the spatio-temporal features, respectively, for emotion classification. The dynamic selection network is responsible for selecting the most suitable stream for every subject. Subject-independent experiments on the benchmarks DEAP, DREAMER, and SEED-IV have readily demonstrated the advantage of DSSN over the related advanced approaches.